Junyu Wang1, Daniel Weller2, Patricia Rodriguez Lozano3, and Michael Salerno1,3,4
1Biomedical Engineering, University of Virginia, Charlottesville, VA, United States, 2Electrical and Computer Engineering, University of Virginia, Charlottesville, VA, United States, 3Medicine, University of Virginia, Charlottesville, VA, United States, 4Radiology, University of Virginia, Charlottesville, VA, United States
Synopsis
First-pass
contrast-enhanced myocardial perfusion imaging is valuable for evaluating coronary
artery disease (CAD). Spiral perfusion imaging
techniques, using a motion-compensated L1-SPIRiT based reconstruction, are
capable of whole-heart high-resolution perfusion imaging. However, this reconstruction
is performed off-line and takes ~1 hour per slice. To address this limitation,
we developed a DEep learning-based Spiral Image REconstruction technique
(DESIRE) for spiral first-pass myocardial perfusion imaging, for both single-slice
(SS) and simultaneous multi-slice (SMS) MB=2 acquisitions, to provide fast and high-quality
image reconstruction and make rapid online reconstruction feasible. High image
quality was demonstrated using the proposed technique for healthy volunteers and
patients.
Introduction
First-pass contrast-enhanced cardiac magnetic resonance
(CMR) perfusion imaging, which is non-invasive and non-radioactive, has proven
to be a valuable tool for evaluating patients with CAD1–4. Recently, we have developed
spiral single-slice (SS) and simultaneous
multi-slice (SMS) perfusion pulse sequence and the motion-compensated
(SMS-Slice)-L1-SPIRiT reconstruction technique capable of whole-heart high-resolution
perfusion imaging5–8. However, this
compressed-sensing based image reconstruction technique is time-consuming and
takes ~1 hour per slice, hence it can’t provide immediate feedback to doctors
and impedes clinical translation. The goal of this study is to develop and
evaluate a DEep
learning-based Spiral Image REconstruction technique (DESIRE) for spiral
first-pass myocardial perfusion imaging, for both SS and SMS MB=2 acquisitions,
to provide fast and high-quality image reconstruction (Figure 1 (a)).Methods
Data Acquisition and Preprocessing SS and SMS
golden-angle spiral perfusion data sets with 1.25×1.25 mm2 in-plane spatial resolution and
whole-heart coverage (6-8 slices) were previously acquired from 18 healthy
volunteers and 4 patients undergoing clinical studies on 3 T SIEMENS
Skyra/Prisma scanners8,9. The in-plane acceleration
factor for SS and SMS MB=2 was approximately 10 and 5, respectively.
Before feeding the data to the network, coil-selection10, motion-correction11, and adaptive phase
combination12 were performed on the
NUFFT-gridded13 multi-coil image series at
each slice location. Images were cropped into
(Frames) to
save memory. Each dynamic image series were normalized to 0-1.
Specifically, for SMS MB=2 acquisitions, to prevent
slice-leakage artifact from being learned by the SMS network, the ground-truth
data were SS L1-SPIRiT images from two slice locations, and the SMS network
input images were the retrospective data from the two separate images with
Hadamard SMS MB=2 phase modulation (Figure 1 (c))14.
156 slices from 20 subjects were used for training, and
another 14 slices from 2 subjects were used for validation. Another 56 slices
from 8 subjects with SS and 76 slices from 10 subjects with SMS MB=2 acquisitions
were used for testing the performance.
Image Reconstruction Network Figure 1 (b)
illustrates the proposed 3D U-Net15 based image reconstruction
network. The inputs to the network were complex-valued under-sampled single-channel
perfusion image series from a given slice location after coil-selection10, motion-correction11, and adaptive phase
combination12. The real and imaginary
values were concatenated into two channels. The outputs were the concatenated
real and imaginary perfusion image series. SS and SMS network were trained
separately where L1-SPIRiT reconstruction results served as the ground truth.
To compare networks, we set the network with
and
without
complex convolution as the baseline. The performance of the network with
respect to the number of kernels (Ninit) at initial layer and the depth of the network (D) was explored (Table 1). Furthermore, we sought to
evaluate complex convolutional networks16 to preserve the phase
information in the raw data to see if this could further improve reconstruction
quality.
Experimental Setup The training of the
baseline network was conducted using PyTorch on a single NVIDIA Tesla P100 GPU
(12 GB memory). The training of the other networks was conducted on four P100 GPUs
due to the memory limitation of a single GPU. All of the trainings were
conducted for 150 epochs with an L1 loss (absolute error) and a batch size of 4.
The shortest training time was baseline network which took ~8 hours, while the
longest training time was complex convolution network which took ~20 hours. All
of the testing experiments were conducted on a single P100 GPU.
Image Analysis Both structural similarity
index (SSIM)17 and root mean square error
(RMSE) for the SS and respective SMS MB=2 reconstructed using DEIRE were
assessed with respect to the ground truth. Prospective SS and SMS images were
graded by an experienced cardiologist (5, excellent; 1, poor).Results
Figure 2 and 3 show examples from the test data for the SS (Figure
2) and SMS MB=2 (Figure 3) reconstructions using the proposed DESIRE technique.
Good image quality was demonstrated.
Figure 4 shows an example case using SS acquisition
undergoing clinical stress spiral perfusion imaging. Good image quality and
temporal fidelity with respect to the ground truth were demonstrated.
Table 1 (a) and (b) show image quality scores for both SS
and SMS MB=2 with different network structures. Table 1 (c) shows the corresponding
reconstruction time per slice of the test data on a NVIDIA Tesla P100 GPU,
while the reconstruction time of using L1-SPIRiT with 80 iterations on an Intel
Xeon CPU (2.40 GHz) was ~1 hour per slice. Increasing the depth and the number of
initial kernels help improve the reconstruction performance. However, for SS
with Ninit=32
, the performance of D=2 is similar to D=3, which indicates the max capacity of the
reconstruction network is reached. Particularly, for the baseline network with
complex convolution operations, the performance is improved. Scores from
cardiologist showed a preference to networks with more initial kernels and those using complex convolutions, which was consistent with the SSIM and RMSE
quantification.Discussion and Conclusion
The proposed image reconstruction network (DESIRE) enabled rapid
and high-quality image reconstruction for both SS and SMS MB=2 whole-heart
ultra-high resolution first-pass spiral perfusion imaging. Further optimization
of the SMS reconstruction network, such as incorporating the through-plane
kernels is still required for optimal performance.Acknowledgements
This work was supported by NIH R01 HL131919 and Wallace H. Coulter
Foundation Grant.
References
1. Greenwood JP, Maredia N,
Younger JF, et al. Cardiovascular magnetic resonance and single-photon emission
computed tomography for diagnosis of coronary heart disease (CE-MARC): a
prospective trial. The Lancet 2012;379:453–460.
2.
Salerno M, Beller GA. Noninvasive Assessment of Myocardial Perfusion. Circ
Cardiovasc Imaging 2009;2:412–424.
3.
Jaarsma C, Leiner T, Bekkers SC, et al. Diagnostic Performance of Noninvasive
Myocardial Perfusion Imaging Using Single-Photon Emission Computed Tomography,
Cardiac Magnetic Resonance, and Positron Emission Tomography Imaging for the
Detection of Obstructive Coronary Artery Disease: A Meta-Analysis. Journal
of the American College of Cardiology 2012;59:1719–1728.
4.
Lipinski MJ, McVey CM, Berger JS, Kramer CM, Salerno M. Prognostic Value of
Stress Cardiac Magnetic Resonance Imaging in Patients With Known or Suspected
Coronary Artery Disease: A Systematic Review and Meta-Analysis. Journal of
the American College of Cardiology 2013;62:826–838.
5.
Yang Y, Kramer CM, Shaw PW, Meyer CH, Salerno M. First-pass myocardial
perfusion imaging with whole-heart coverage using L1-SPIRiT accelerated
variable density spiral trajectories: Whole Heart Spiral Myocardial Perfusion
Imaging. Magnetic Resonance in Medicine 2016;76:1375–1387.
6.
Yang Y, Meyer CH, Epstein FH, Kramer CM, Salerno M. Whole‐heart spiral
simultaneous multi‐slice first‐pass myocardial perfusion imaging. Magnetic
Resonance in Medicine 2019;81:852–862.
7.
Wang J, Yang Y, Zhou R, Jacob M, Weller DS, Salerno M. SMS Slice L1-SPIRiT:
auto-calibrated image reconstruction for spiral simultaneous multi-slice
first-pass perfusion imaging with 1.25 mm resolution and whole heart coverage
at 3T. In: Proceedings of the SCMR 23rd Annual Scientific Sessions, Orlando,
Florida, USA, 2020.
8.
Wang J, Yang Y, Zhou R, et al. High resolution spiral simultaneous multi‐slice
first‐pass perfusion imaging with whole-heart coverage at 1.5 T and 3 T. In:
Proceedings of the 28th Annual Meeting of ISMRM, 2020. p 1313.
9.
Yang Y, Kramer C, Salerno M. Whole heart First-pass spiral perfusion imaging
with 1.25mm resolution at 3T. In: Proceedings of the 26th Annual Meeting of
ISMRM, Paris, France, 2018. p 3321.
10.
Zhou R, Yang Y, Mathew RC, et al. Free-breathing cine imaging with
motion-corrected reconstruction at 3T using SPiral Acquisition with Respiratory
correction and Cardiac Self-gating (SPARCS). Magnetic Resonance in Medicine
2019;82:706–720.
11.
Zhou R, Huang W, Yang Y, et al. Simple motion correction strategy reduces respiratory-induced
motion artifacts for k-t accelerated and compressed-sensing cardiovascular
magnetic resonance perfusion imaging. J Cardiovasc Magn Reson 2018;20:6.
12.
Walsh DO, Gmitro AF, Marcellin MW. Adaptive reconstruction of phased array MR
imagery. Magnetic Resonance in Medicine 2000;43:682–690.
13.
Fessler JA, Sutton BP. Nonuniform fast Fourier transforms using min-max
interpolation. IEEE Transactions on Signal Processing 2003;51:560–574.
14.
Souza SP, Szumowski J, Dumoulin CL, Plewes DP, Glover G. SIMA: simultaneous
multislice acquisition of MR images by Hadamard-encoded excitation. J Comput
Assist Tomogr 1988;12:1026–1030.
15.
Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical
Image Segmentation. arXiv:1505.04597 [cs] 2015.
16.
Cole E, Cheng J, Pauly J, Vasanawala S. Analysis of Deep Complex-Valued
Convolutional Neural Networks for MRI Reconstruction. :10.
17.
Zhou Wang, Bovik AC, Sheikh HR, Simoncelli EP. Image quality assessment: from
error visibility to structural similarity. IEEE Transactions on Image
Processing 2004;13:600–612.